List your company

Strategic Reasoning on HardCore Don'tSayIt OOD (held-out variant)

22.92Win RateDEPT
-0.51125.571911.65517.7381May 9, 2026

Evaluation results

MethodLinks
DEPTBackbone=Qwen3-4B-Base2026.0522.92
GRPOBackbone=Qwen3-4B-Base2026.0522.01
DEPTBackbone=Qwen3-8B-Base2026.0519.27
SPIRALBackbone=Qwen3-8B-Base2026.0518.88
SPIRALBackbone=Qwen3-4B-Base2026.0517.97
SPAGBackbone=Qwen3-8B-Base2026.0512.12
MARSBackbone=Qwen3-4B-Base2026.0512.11
GRPOBackbone=Qwen3-8B-Base2026.0510.03
MARSBackbone=Qwen3-8B-Base2026.058.33
SPAGBackbone=Qwen3-4B-Base2026.057.03
VANILLABackbone=Qwen3-8B-Base2026.052.34
VANILLABackbone=Qwen3-4B-Base2026.050.39